Signalling a vertex is like adding it to the work queue. What this slide is saying is that instead of just signalling all of the vertices once in each iteration (which acts like an implicit barrier), we work on some extreme cases more often by signalling them more than once per iteration, which adds them to the dynamic work queue before the next full iteration over all the vertices.
This comment was marked helpful 0 times.
akashr
In this example, when the vertex gets added to the dynamic work queue, there is no specific time when it will get removed from the dynamic work queue. So when the vertex gets removed the worker queue, its neighbors data might not be the latest. Those neighbors might be in the midst of getting their data updated. We would have to do ensure consistency which is not done here if we want to make sure we always have updated data for each and every neighbor.
Signalling a vertex is like adding it to the work queue. What this slide is saying is that instead of just signalling all of the vertices once in each iteration (which acts like an implicit barrier), we work on some extreme cases more often by signalling them more than once per iteration, which adds them to the dynamic work queue before the next full iteration over all the vertices.
This comment was marked helpful 0 times.
In this example, when the vertex gets added to the dynamic work queue, there is no specific time when it will get removed from the dynamic work queue. So when the vertex gets removed the worker queue, its neighbors data might not be the latest. Those neighbors might be in the midst of getting their data updated. We would have to do ensure consistency which is not done here if we want to make sure we always have updated data for each and every neighbor.
This comment was marked helpful 0 times.